Review Model implementation

          # **Task 2 - Model Eos30f3 Reproducibility**

Identify Results you want to reproduce

From what can be seen from the publication, various classification models were trained using a neural network called directed message passing neural network (D-MPNN), with diverse Datasets obtained from multiple sources to identify compounds that block hERG.

The Model with the optimal performance is D-MPNN + moe206 with an AUC-ROC value of 0.956 ± 0.005. But the Molecular descriptor (moe206) that was used for this is closed-source, so the model that was implemented on Ersilia was trained without a molecule Featurizer.

The original D-MPNN model with an AUC-ROC value of (AUC-ROC 0.947 ± 0.005) is what we are going to Reproduce, which was achieved on a random splitting using a 5-fold cross validation, it was trained on a dataset of 7889 compounds with well-defined experimental data on the hERG and with diverse chemical structures and has 6 thresholds (10 μM, 20 μM, 40 μM, 60 μM, 80 μM, and 100 μM) for distinguishing hERG blockers from non-blockers. The Author chooses a 10 μM threshold for the model.

This dataset was assembled by Cai et al in their work published in J Chem Inf Model, 2019. The Datasets is available here

Implement the model on your system as described by the authors

I cloned the model repository into my Ubuntu 22.4 system using git clone https://github.com/AI-amateur/DMPNN-hERG.git

The model requires Chemprop Installation, so I navigated to the path where Chemprop is.

cd dmpnn-herg/chemprop

Created the virtual environment and installed the packages and dependencies in the environment.yml file then proceed to activate the environment.

conda env create -f environment.yml
conda activate chemprop

Check that the Model provides the same result as Eos30f3 on EMH

The AUC-ROC value of 0.947 ± 0.005 I’m trying to reproduce was gotten on the test sets, as seen here

To reproduce the AUC-ROC value since I don't have access to the test datasets.

I modified the reproduced shell script from here to train and test the model without the Featurizer :

python chemprop/train.py --data_path Figure2/Cai_TableS3_fixed.csv --dataset_type classification --save_dir Figure2/result --epochs 50 --split_type random  --num_folds 5 --quiet

And the AUC-ROC value of (AUC-ROC 0.947 ± 0.006) was reproduced successfully.

result.log

results

Note: The eos30f3 Model that was implemented on Ersilia is not the DPMN model that was trained with cai datasets of 7889 compounds, but one that was implemented from this checkpoints here.

It was trained on 393 Compounds as seen here

To further confirm this, I ran predictions with the test datasets provided, on the models in the checkpoint:

python chemprop/predict.py \
--checkpoint_dir example_notebook/I_train_rand \
--test_path example_notebook/I_test.csv \
 --preds_path DPMN_preds.csv

The same test datasets on the eos30f3 model on Ersilia Model Hub and obtained the same values across all classification metrics.

Models	Accuracy	Precision	Recall	F1 Score	ROC AUC	Balanced Accuracy	MCC	Cohen's Kappa
eos30f3	0.7897	0.8014	0.907	0.8509	0.7929	0.7338	0.5101	0.499
DPMN-hERG	0.7897	0.8014	0.907	0.8509	0.7929	0.7338	0.5101	0.499

Originally posted by @Malikbadmus in https://github.com/ersilia-os/ersilia/issues/979#issuecomment-1996008631

ersilia-os / eos30f3